Automatic segmentation of large-scale CT image datasets for detailed body composition analysis

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiology

Background:

• Body composition (BC) is crucial for assessing risks of type 2 diabetes and cardiovascular disease.
• Manual segmentation of CT images for BC analysis is time-consuming and subjective.
• Automated techniques are needed for efficient and objective BC assessment.

Purpose of the Study:

• To develop and evaluate fully automated segmentation techniques for body composition analysis.
• To apply these techniques to a 3-slice CT imaging protocol (liver, abdomen, thigh).
• To enable detailed analysis of numerous tissues and organs relevant to BC.

Main Methods:

• Trained and evaluated four convolutional neural network architectures (ResUNET, UNET++, Ghost-UNET, Ghost-UNET++) on over 4000 CT subjects.
• Segmented multiple tissues: liver, spleen, muscle, bone marrow, cortical bone, and various adipose depots (VAT, IPAT, RPAT, SAT, DSAT, SSAT, IMAT).
• Utilized tenfold cross-validation and test sets for model validation.

Main Results:

• UNET++ achieved the highest performance with a Dice score of 0.981.
• Ghost-UNET++ demonstrated competitive results and superior computational efficiency.
• All models showed strong performance in segmenting various body composition components.

Conclusions:

• Fully automated segmentation is effective for analyzing multiple tissues and organs from 3-slice CT scans for BC.
• UNET++ and Ghost-UNET++ offer accurate and efficient automated segmentation solutions.
• Automated methods significantly reduce analysis time and enhance objectivity in large-scale BC studies.